Distributed frameworks for edgeAI span the edge to cloud continuum in support of AI through services for deployment, monitoring, and optimization. However, edge networks are unreliable, raising monitoring issues including missing events, network interruptions, and storage limitations all of which can lead to information loss that undermines decision-making and analytics. We address information loss through Patra Relational Graph Convolutional Networks (Patra-RGCN), an extension of Relational Graph Convolutional Networks (R-GCN), that uses both structural and content information about a graph (including time and Model Card subgraphs) to infer missing edges and suggest new connections. Experimental evaluation demonstrates that by effectively capturing both structural and attribute-level information, the proposed model significantly improves the detection of information loss in heterogeneous graphs.

This paper introduces a framework for Model Cards, Patra ModelCards, that embeds model cards in the edge-cloud continuum for semi-automated information capture with the objective of greater trustworthiness and accountability for AI/ML models. Information captured includes fairness, explainability, and behavior of a model in different deployed environments. Our evaluation is of the framework’s claim of greater accountability. We evaluate the use of embedded vectors and similarity analysis to distinguish between deployed models that are duplicates of each other from those that represent a revision of an earlier developed model. The evaluation shows promising outcomes and good performance.

The edge-cloud-HPC continuum is transformative for AI processing at the edge. With greater availability of both edge and cloud resources, AI inference, training, and optimization can be distributed across the continuum. We target edge-cloud in particular where the workload at the Edge server can exhibit discrete changes, for instance, when motion is detected. We optimize for Quality of Experience (QoE) and utilize historical data from the Edge, graphs, and Deep Learning to infer the next action to take. Using a large synthetic workload and publicly profiled inference models, our results show that predictive guidance outperforms random choice or best guess in optimal QoE of the edge-cloud continuum.

The campaign is an experimentation construct for codesign activity wherein multiple researchers carry out computational experiments that individually contribute to a shared goal. The larger objective of our research is a system that exists in the experimental environment that constructs a knowledge representation of campaigns and products both produced and consumed such that the campaign can as efficient as possible and the products richly contextualized for reuse. Using campaign experiments running on the Summit machine at Oak Ridge National Labs, we demonstrate early results of support for discovery queries and for detecting when two sweeps are similar.

In the business and research landscape of today, data analysis consumes public and proprietary data from numerous sources, and utilizes any one or more of popular data-parallel frameworks such as Hadoop, Spark and Flink. In the Data Lake setting these frameworks co-exist. Our earlier work has shown that data provenance in Data Lakes can aid with both traceability and management. The sheer volume of fine-grained provenance generated in a multi-framework application motivates the need for on-the-fly provenance processing. We introduce a new parallel stream processing algorithm that reduces fine-grained provenance while preserving backward and forward provenance. The algorithm is resilient to provenance events arriving out-of-order. It is evaluated using several strategies for partitioning a provenance stream. The evaluation shows that the parallel algorithm performs well in processing out-of-order provenance streams, with good scalability and accuracy.